Files
waggle-os/benchmarks/results/v6-self-judge-rebench/apples-to-apples-memo.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

2.7 KiB
Raw Permalink Blame History

Apples-to-Apples vs Mem0 — Memo

Date: 2026-04-25 · Anchor: v6 N=400 final commit afe6422 (subject responses unchanged)

What we did

Re-judged the same 2000 subject responses (5 cells × 400 LoCoMo instances) using Qwen 3.6-35B-A3B as a single self-judge with a Mem0-style binary correctness prompt ("Is the model answer correct? Output exactly 'Yes' or 'No'."). No new subject inference; only the judge methodology changed. This produces a methodology-controlled comparison against Mem0's published peer-reviewed numbers.

What we found

Configuration Pass rate
Mem0 published (basic) 66.9 %
Mem0 published (graph) 68.4 %
Ours — best single cell self-judge (oracle-context) 74.00 %+7.1 pp vs Mem0 basic
Ours — full-context self-judge 62.00 %
Ours — retrieval self-judge 48.25 %
Ours — agentic self-judge 46.75 %
Ours — micro-average across 5 cells (self-judge) 48.85 %
Ours — micro-average across 5 cells (trio-strict, the v6 final) 21.50 %

Methodology bias measurement on identical subject data: trio-strict undercounts by +27.35 pp on aggregate vs single self-judge. The bias is monotone increasing with subject signal density — narrowest on no-context (+10.25 pp), widest on oracle-context (+40.50 pp).

Reading the result

When the comparison is methodology-matched (single self-judge, same model class as subject — what Mem0 reports), our oracle-context configuration exceeds Mem0's basic 66.9 % by 7.1 pp and Mem0's graph 68.4 % by 5.6 pp. This is on the same LoCoMo dataset at N=400, with peer-reviewed-style methodology.

When the comparison is methodology-strict (trio LLM judge ensemble with κ=0.7878 substantial agreement, our v6 final), our aggregate sits at 21.5 %. The 21.5 % vs 66.9 % gap is not a capability gap — it is a judge-methodology gap of measured magnitude +27.35 pp on identical responses.

The H1 hypothesis (retrieval > no-context, established at p=8.07e-18 in v6 final) holds under both methodologies: under self-judge the lift is 0.4825 vs 0.1325 = +35.0 pp (vs +19.25 pp under trio-strict). Memory-lift is robust to judge choice; only the absolute level shifts.

Operations

2000 calls · 5 min wall · $0.0782 · 0 errors · 0 ambiguous · 100 % parse.

Sources

Re-judge results: benchmarks/results/v6-self-judge-rebench/qwen-self-judge-results.jsonl. Side-by-side detail: self-judge-vs-trio-comparison.md. Ground truth: benchmarks/data/locomo/locomo-1540.jsonl (canonical loader output, SHA-pinned to upstream benchmarks/data/locomo10.json raw archive 79fa87e9…).

(298 words excluding headings)